Hi Anthony,let me throw a few random problems without really thinking this through. First of all, the whole thing relies on HashMaps, but we can't really use nodes in hash-based collections as they are not immutable. To maintain the registrations with the rectangles, we would need to do some computations (bounds) eagerly. There is again the problem of restricting to touch-sensitive nodes only. Finally, even if your proposal turns out to work like a charm, it reduces the number of examined nodes, but the most important and difficult part - how do we choose the right one - needs to be specified.
On 20.11.2013 10:02, Anthony Petrov wrote:
How about you divide the top-level window surface on equal rectangles of size, say, 0.5"x0.5" (the size in pixels will depend on the screen DPI.) Each rectangle is basically a hashmap of nodes.The nodes that are touch sensible register and update themselves with the rectangles when their layout information changes (when they're moved and/or resized). This should work reasonably fast given O(1) complexity for hashmap operations.When processing a touch event, given the low-resolution of the rectangle mesh, we can quickly identify a set of rectangles that a touch area (an oval) intersects. There usually will be no more than four such rectangles. After that we construct a union of all nodes laying in those rectangles, and simply iterate over the union to choose a node laying closest to the touch area (or having the largest intersection with it, whatever.) Again, this should work fast because fetching nodes from the hashmaps is fast, and the number of nodes that need to be examined should be relatively small.-- best regards, Anthony On 11/19/2013 08:18 PM, Daniel Blaukopf wrote:On Nov 19, 2013, at 4:34 PM, Pavel Safrata <pavel.safr...@oracle.com> wrote:Hello Daniel, On 17.11.2013 15:09, Daniel Blaukopf wrote:Hi Pavel,I think we we do use CSS to configure feel as well as look - and this is feel, not look - but I also don’t feel strongly about whether this needs to be in CSS.This was exactly my point.Yes, we agree. I should have phrased the above differently to show that.For now it certainly is sufficient. However these touch devices can describe more complex shapes for the touch contact area and it is possible we will want to take advantage of that in the future. We should certainly optimize for a circle or oval shape.I like your idea of simply picking the closest touch sensitive node that is within range. That puts the burden on the touch event to describe what region it covers. On the touch screens we are currently looking at, a region would be defined as an oval - a combination of centre point, X diameter and Y diameter. However the touch region can be any shape, so might need to be represented as a Path.I don't think we need to be precise about the shape, all of this is about fixing the imprecise touch where the shape is rather accidental, I think even a circle with an average diameter would be sufficient to achieve the goal.I realize that it is possible to describe touch sensitive concave shapes, but I am not sure they matter for this. If developers are going to go to the trouble of defining a concave shape that they want to be touch sensitive within its area but not in all of its bounding box, are they really then going to want that area to be extended? I’d consider a concave touch shape with extended capture zone to be sufficiently unlikely that we could treat it as concave. Which, I realize is not quite what my proposed algorithm does.Iterating over pixels just isn’t going to work though. If we have a 300dpi display the touch region could be 150 pixels across and have an area of nearly 18000 pixels. Instead we’d want a way to ask a parent node, “In your node hierarchy, which of your nodes’ borders is closest to this region”. So we’d need to come up with an efficient algorithm to answer this question. We’d only ask this question for nodes with extended capture zone.The nodes with extended capture zone may be fully or partially hidden behind nodes without it and this needs to be taken into account. How would it be done when the question would be limited to nodes with extended capture zone? And even if it wasn't, what would the position of the border tell us about the area covered by the node? It can be a line.. Also what I wanted to achieve is pick the node which has visible pixels in the picking area even if its border is hidden behind other node or is somewhere far away. This wouldn't be possible, would it?We could reasonably limit the algorithm to dealing with convex shapes.Can we? What about paths, polygons etc?You are correct, my algorithm won’t find the closest node. Back to the drawing board.Then we can consider an imaginary line L from the node center point to the touch center point. The intersection of L with the node perimeter is the closest point of contact.It isn't. Imagine a very wide and flat (small height) rectangle and the touch point directly above its upper-left corner. The closest point is the corner, not the point on L which is close to rectangle's center and may be many times farther.Thanks, DanielIf this point is also within the touch area then we have a potential match. We iterate over all nearby nodes with extended capture zone in order to find the best match.This will then be O(n) in both time and space for n nearby nodes, given constant time to find the intersection of L with the node perimeter. This assumption will be true for rectangular, oval and rounded rectangle nodes.So in summary, if I understand this algorithm correctly I don't think it's going to work. On the other hand, I admit that computing 18000 pixels is probably not viable. Right now I don't have any solution, I'll continue thinking..Thanks, PavelThanks, DanielOn Nov 15, 2013, at 11:09 PM, Pavel Safrata <pavel.safr...@oracle.com> wrote:Hello, let me start with a few comments."changing behavior based on which nodes have listeners on them" - absolutely not. We have capturing, bubbling, hierarchical event types, so we can't decide which nodes listen (in the extreme case, scene can handle Event.ANY and perform actions on the target node based on the event type)."position does not fall in the boundaries of the node" - I don't think it will be very harmful. Of course it's possible for users to write handlers that will be affected, but I don't think it happens often, it seems quite hard to invent such handler. The delivery mechanism should be absolutely fine with it, we have other cases like that (for instance, dragging can be delivered to a node completely out of mouse position). Of course picking a 3D node in its capture zone would mean useless PickResult (texture coordinates etc.)CSS-accessible vs. property-only - I don't have a strong opinion. I agree it's rather "feel" than "look", on the other hand I think there are such things already (scrollbar policy for instance).Now I'll bring another problem to the concept. Take the situation from Daniel's original picture with two siblings competing for the capture zones:The following may be a little wild, but anyway - have one of the siblings with capture zone and the other one without it, the one without it partly covering the one with it. Wouldn't it be great if the capture zone was present around the visible part of the node (reaching over the edge of the upper node)? I think it would be really intuitive (fuzzy picking of what you see), but it's getting pretty complicated.From now on, I'll call the node with enabled capture zone "touch sensitive".The only algorithm I can think of that would provide great results is: - Pick normally at the center. If the picked node is touch sensitive, return it. - Otherwise, run picking for each pixel in the touch area, find the closest one belonging to a touch sensitive node and return that node (if there is none, then of course return the node at the center).Obviously we can hardly do so many picking rounds. But it can be significantly optimized: - Perform the area picking in one pass, filling an array - representing pixels - by the nodes picked on them- Descend only when bounds intersect with the picking area- Don't look farther from the center than the already found best match- Don't look at pixels with already picked node- For many nodes (rectangular, circular, with pickOnBounds etc.), instead of testing containment many times, we can quickly tell the intersection with the picking area- Perhaps also checking each nth pixel would be sufficientThis algorithm should be reasonably easy to code and very robust (not suffering from various node-arrangement corner-cases), but I'm still not sure about the performance (depends mostly on the capture zone size - 30-pixel zones may result in calling contains() nearly thousand times which might kill it). But perhaps (hopefully) it can be perfected. Right now I can't see any other algorithm that would work well and would result in more efficient implementation (the search for overlapping nodes and closest borders etc. is going to be pretty complicated as well, if it's even possible to make it work).What do you think? Any better ideas? Pavel On 13.11.2013 22:09, Daniel Blaukopf wrote:Hi Seeon, Summarizing our face to face talk today:I see that the case described by Pavel is indeed a problem and agree with you that not every node needs to be a participant in the competition for which grabs touch input. However I’m not keen on the idea of changing behavior based on which nodes have listeners on them. CSS seems like the place to do this (as I think Pavel suggested earlier). In Pavel’s case, either: - the upper child node has the CSS tag saying “enable extended capture zone” and the lower child doesn’t: then the upper child’s capture zone will extend over the lower child - or both will have the CSS tag, in which case the upper child’s capture zone would be competing with the lower child’s capture zone. As in any other competition between capture zones the nearest node should win. The effect would be the same as if the regular matching rules were applied on the upper child. It would also be the same as if only the lower child had an extended capture zone. However, I’d consider this case to be bad UI programming.We agreed that “in a competition between capture zones, pick the node whose border is nearest the touch point” was a reasonable way to resolve things.Thanks, DanielOn Nov 13, 2013, at 12:31 PM, Seeon Birger <seeon.bir...@oracle.com> wrote:Hi Pavel,Your example of 'child over child' is an interesting case which raises some design aspects of the desired picking algorithm: 1. Which node to pick when one node has a 'strict containership' over the touch center and the other node only has a fuzzy containership (the position falls in the fuzzy area).2. Accounting for z-order for extended capture zone area. 3. Accounting for parent-child relationship. Referring to your 'child over child' example: http://i.imgur.com/e92qEJA.jpgThe conflict would arise were touch point center position falls in the capture zone area of child2 but also clearly falls in the strict bounds of child1. Generally, when two control nodes compete on same touch event (e.g. child1 & child2 in Daniel's diagram), it seems that we would like to give priority to "strict containership" over "fuzzy containership".But in your case it's probably not the desired behavior.Also note that in the general case there's almost always exists come container/background node that strictly contains the touch point, but it would probably be an ancestor of the child node, so the usual parent-child relationship order will give preference to the child.One way out it is to honor the usual z-order for the extended area of child2, so when a touch center hits the fuzzy area of child2, then child2 would be picked.But is not ideal for Daniel's example: http://i.imgur.com/ELWamYp.pngwhere the 2 nodes don't strictly overlap, but their capture zones do. Preferring one child by z-order (which matches the order of children in the parent) is not natural here. And we might better choose the node which is "closer"To the touch point. So to summarize I suggest this rough picking algorithm:1. Choose all uppermost nodes which are not transparent to mouse events and contain the touch point center either strictly or by their capture zone. 2. Remove all nodes that is strictly overlapped by another node and is below that node by z-order. 3. Out of those left choose the "closest" node. (the concept of "closet" should employ some calculation which might not be trivial in the general case). 4. Once a node has been picked, we follow the usual node chain list for event processing.Care must be taken so we not break the current model for event processing. For example, if a node is picked by its capture zone, it means that the position does not fall in the boundaries of the node, so existing event handling code that relies on that would break. So I think the capture zone feature should be selectively enabled for certain type of nodes such buttons or other classic controls.Regards, Seeon -----Original Message----- From: Pavel Safrata Sent: Tuesday, November 12, 2013 1:11 PM To: Daniel Blaukopf Cc: OpenJFX Subject: Re: discussion about touch events (Now my answer using external link) Hello Daniel,this is quite similar to my idea described earlier. The major difference is the "fair division of capture zones" among siblings. It's an interesting idea, let's explore it. What pops first is that children can also overlap. So I think it would behave like this (green capture zonesomitted):Child in parent vs. Child over child: http://i.imgur.com/e92qEJA.jpg..wouldn't it? From user's point of view this seems confusing, both cases look the same but behave differently. Note that in the case on the right, the parent may be still the same, developer only adds a fancy background as a new child and suddenly the red child can't be hit that easily. What do you think? Is it an issue? Or would it not behave this way?Regards, Pavel On 12.11.2013 12:06, Daniel Blaukopf wrote:(My original message didn't get through to openjfx-dev because I usedinline images. I've replaced those images with external links)On Nov 11, 2013, at 11:30 PM, Pavel Safrata <pavel.safr...@oracle.com<mailto:pavel.safr...@oracle.com>> wrote:On 11.11.2013 17:49, Tomas Mikula wrote:On Mon, Nov 11, 2013 at 1:28 PM, Philipp Dörfler <phdoerf...@gmail.com <mailto:phdoerf...@gmail.com>> wrote:I see the need to be aware of the area that is covered by fingersWhat about picking the one that is closest to the center of the touch?rather than just considering that area's center point. I'd guess that this adds a new layer of complexity, though. For instance:Say we have a button on some background and both the background andthe button do have an onClick listener attached. If you tap thebutton in a way that the touched area's center point is outside of the buttons boundaries - what event will be fired? Will both the background and the button receive a click event? Or just either thebackground or the button exclusively? Will there be a new event type which gets fired in case of such area-based taps? My suggestion would therefore be to have an additional area tapevent which gives precise information about diameter and center of the tap. Besides that there should be some kind of "priority" forchoosing which node's onClick will be called.There is always something directly on the center of the touch(possibly the scene background, but it can have event handlers too).That's what we pick right now. PavelWhat Seeon, Assaf and I discussed earlier was building some fuzzinessinto the node picker so that instead of each node capturing only events directly on top of it: Non-fuzzy picker: http://i.imgur.com/uszql8V.png ..nodes at each level of the hierarchy would capture events beyond their borders as well: Fuzzy picker: http://i.imgur.com/ELWamYp.png In the above, "Parent" would capture touch events within a certainradius around it, as would its children "Child 1" and "Child 2". Since "Child 1" and "Child 2" are peers, they would have a sharp division between them, a watershed on either side of which events would go to one child node or the other. This would also apply if the peer nodes were further apart; they would divide the no-man's land between them. Of course this no-man's land would be part of "Parent" and could could be touch-sensitive - but we won't consider "Parent" as an event target until we have ruled out using one of its children's extended capturezones.The capture radius could either be a styleable property on the nodes,or could be determined by the X and Y size of a touch point asreported by the touch screen. We'd still be reporting a touch point, not a touch area. The touch target would be, as now, a single node.This would get us more reliable touch capture at leaf nodes of the node hierarchy at the expense of it being harder to tap the background. This is likely to be a good trade-off. DanielTomasMaybe the draw order / order in the scene graph / z buffer valuemight be sufficient to model what would happen in the real, physical world.Am 11.11.2013 13:05 schrieb "Assaf Yavnai" <assaf.yav...@oracle.comThe ascii sketch looked fine on my screen before I sent the mail :( I hope the idea is clear from the text (now in the reply dialogits also look good) Assaf On 11/11/2013 12:51 PM, Assaf Yavnai wrote:Hi Guys,I hope that I'm right about this, but it seems that touch events in glass are translated (and reported) as a single point events(x & y) without an area, like pointer events. AFAIK, the controls response for touch events same as mouseevents (using the same pickers) and as a result a button press, for example, will only triggered if the x & y of the touch eventis within the control area. This means that small controls, or even quite large controls(like buttons with text) will often get missed because the 'strict'node picking, although from a UX point of view it is strange as the userclearly pressed on a node (the finger was clearly above it) butnothing happens... With current implementation its hard to use small features incontrols, like scrollbars in lists, and it almost impossible to implement something like 'screen navigator' (the series of small dots in the bottom of a smart phones screen which allow you tojump directly to a 'far away' screen)To illustrate it consider the bellow low resolution sketch, wherethe "+"is the actual x,y reported, the ellipse is the finger touch areaand the rectangle is the node.With current implementation this type of tap will not trigger thenode handlers __ / \ / \___/ __+_ \___ in this scenario the 'button' will not getpressed | \ / | |___\ ___ / __ | \___/ If your smart phone support it, turn on the touch debuggingoptions in settings and see that each point translate to a quite large circle and what ever fall in it, or reasonably close to it,get picked.I want to start a discussion to understand if my perspective isaccurate and to understand what can be done, if any, for the coming release or the next one.We might use recently opened RT-34136 <https://javafx-jira.kenai. com/browse/RT-34136> for logging this, or open a new JIRA for itThanks, Assaf